In 2021, 17,269 residents of Maine died, and 6,208 residents of Alaska. The two numbers say how much work the two states' hospitals, hospices and funeral homes had that year. They cannot say where the risk of dying was higher, because Maine had nearly twice as many people. Dividing each count by its population gives 1,258.4 deaths per 100,000 residents in Maine and 847.3 in Alaska. Was a Maine resident's chance of dying that year really about half again an Alaskan's?
These figures are crude death rates: all of a population's deaths in a year divided by all of its residents. They appear to settle the question, and the conclusion they suggest is wrong. The reason it is wrong is the subject of this unit, Rates and Standardization.
This article builds the measures themselves. First, we separate rates from proportions and ratios, and count the time people spend at risk. Then we distinguish incidence from prevalence, and assemble a crude rate from the rates of the age groups within it. Finally we measure how much random variation a rate based on a given number of deaths carries, and compare the two states. By the end, the comparison of Maine and Alaska will be seen to be statistically precise and still misleading. The adjustment that corrects it is the subject of Direct and Indirect Standardization.
The article uses three sources of data:
- the final 2021 death counts published by the National Center for Health Statistics (NCHS) in Deaths: Final Data for 2021 (Murphy et al., 2024);
- the Census Bureau's population estimates for July 1, 2021 by single year of age, which NCHS used as its denominators;
- deaths of U.S. residents by age from CDC WONDER, the federal online query system for vital statistics.
Every number below is computed by a script in the site's repository. Where NCHS published the same figure, the script checks it.
Counts, proportions, ratios and rates
Before we can compare the two states, we need to know what kind of number a "death rate" is. Three kinds of fraction appear in epidemiology, and the words for them are used loosely, sometimes by the agencies that publish them. They differ in what the denominator contains.
A ratio is any quotient of two quantities, with no requirement that the numerator be part of the denominator. The sex ratio is the standard example: on July 1, 2021, Maine had 676,493 male and 695,754 female residents, 97.2 males per 100 females. Alaska had 384,204 and 348,469, or 110.3. The numerator and denominator are separate groups, and the ratio can take any positive value.
A proportion is a ratio whose numerator is a subset of its denominator: the share of a population with some characteristic. It is dimensionless and lies between 0 and 1. In 2021, 21.7% of Maine's residents were 65 or over, against 13.3% of Alaska's. The share of adults with hypertension, estimated from NHANES in Inference on Proportions, is a proportion. So is the share of a cohort that develops a disease within five years.
A rate, in the strict sense used here, measures how fast events occur. It is the number of events divided by the total amount of time the population spent at risk of them. Its denominator is not a number of people but a quantity of person-time. So a rate has units of events per person per unit of time, for example deaths per 100,000 person-years. A rate is not bounded by 1. Its value depends on the unit of time: 50 deaths per 1,000 person-years is the same rate as 4.2 per 1,000 person-months.
The terms blur in practice. A "death rate" of 1,258.4 per 100,000 is a rate in the strict sense. Its denominator, the midyear population, stands in for the person-years lived during the year (see below). The infant mortality rate divides infant deaths by live births in the same year, a ratio of two counts rather than a rate. Reading a published figure correctly starts with asking what its denominator contains, and over what time.
Check your understanding. A "case-fatality rate" is the share of people with a disease who die of it. Is it a proportion, a ratio or a rate? What about a "prevalence rate", the share of a population who have a disease?
Both are proportions, despite their names. Each numerator (the people who died, the people who are ill) is part of its denominator, and neither denominator contains any time. The same goes for an "attack rate", the share of an exposed group who fall ill during an outbreak.
Every rate also has a scale, a power of 10 chosen so that the values are readable: per 100 (percent), per 1,000 or per 100,000 people per year. The scale is cosmetic, but rates compared with each other must share it, along with the definitions of the event and of the population at risk.
That leaves the denominator of a rate: time at risk. How is it counted when people are watched for different lengths of time?
Person-time
In a cohort study, people enter and leave observation at different times. Some join after the study opens, some are lost to follow-up, and some die of other causes. Those still under observation when the study closes are followed no further. A person watched for six months cannot contribute as much information about the rate of an event as a person watched for five years.
The solution is to count person-time. Each person contributes the time from their entry until they have the event, leave observation or the study ends, whichever comes first. The denominator of the rate is the sum of these contributions.
The figure shows a small cohort of our own construction.
- Persons 2 and 6 develop the disease, after 2.5 and 3.5 years of observation.
- Person 4 is lost after 3 years, and person 7 dies of another cause after 1.5 years.
- Persons 3, 5, 6 and 8 join after the study opens.
The person-years add up to
$$5.0 + 2.5 + 4.0 + 3.0 + 3.0 + 3.5 + 1.5 + 2.0 = 24.5,$$
and the incidence rate is
$$\hat\lambda = \frac{\text{new cases}}{\text{person-time at risk}} = \frac{2}{24.5\ \text{person-years}} = 0.0816 \text{ per person-year},$$
or 81.6 cases per 1,000 person-years.
Why not simply report 2 cases among 8 people? The naive proportion 2/8 = 0.25 answers no well-defined question here, because the eight people were not all watched for the same length of time. It is neither a five-year risk, since several people were followed for much less, nor anything per year.
Person-time counts only time at risk. Once person 2 has the disease, the time after diagnosis is not added, because that person can no longer become a new case. For a study of deaths from all causes, everyone alive is at risk, and the distinction disappears.
From a rate to a risk
A rate says how fast cases occur, but a patient usually wants something else: the chance of becoming a case within some period. The two are different quantities, but they are related.
Here is the idea. If cases occur at a constant rate among those still free of the disease, the pool of people who can become cases shrinks as they do. So the risk grows more slowly than the rate times the time.
The formal version. Suppose the rate $\lambda$ is constant over time. Among people still free of the disease at any moment, cases then occur at $\lambda$ per person-year. Let $S(t)$ be the proportion still free of disease after time $t$. It falls at a rate proportional to itself, $S'(t) = -\lambda S(t)$, with $S(0) = 1$, and the solution is $S(t) = e^{-\lambda t}$. The risk (cumulative incidence) over time $t$ is the probability of becoming a case by then:
$$\text{risk}(t) = 1 - e^{-\lambda t} \approx \lambda t \quad \text{when } \lambda t \text{ is small}.$$
Breslow and Day (1987, §2.2) use the same conversion, $P = 1 - \exp(-\Lambda)$, to turn cumulative cancer rates into lifetime risks.
Check your understanding. The hypothetical cohort's rate is 0.0816 per person-year. Is the five-year risk above or below $5 \times 0.0816 = 0.41$?
Below: the five-year risk is $1 - e^{-5 \times 0.0816} = 0.335$, because people who have already become cases cannot become cases again. Over one year the shrinkage hardly matters: the one-year risk is $1 - e^{-0.0816} = 0.0784$, close to the rate itself.
The formula ignores competing causes of death, which remove people from risk without making them cases. With a common competing risk, the actual proportion of a cohort that becomes cases is lower.
The midyear population as person-time
Back to the two states. Vital statistics do not follow individuals, so where is the person-time in a state's death rate? A state's death rate for a calendar year divides that year's deaths by the population on July 1. Suppose the population changes roughly linearly during the year. Then its average over the year is close to its midyear value. That average is exactly the person-years lived divided by one year. So the midyear population is an estimate of person-years, and the annual death rate is a rate in the strict sense. NCHS computes rates for years other than census years with populations "estimated as of July 1" (Murphy et al., 2024, Technical Notes).
Incidence and prevalence
A death happens once, but a disease has a beginning and a duration, and that gives two different ways to say how common it is. They answer different questions.
- Incidence measures new cases. The incidence rate is the number of new cases per unit of person-time among people free of the disease. The cumulative incidence (risk) is the proportion of an initially disease-free group who become cases within a stated period. Incidence describes the flow into the diseased state and is the measure for studying causes, since an exposure that raises the risk of disease raises the incidence.
- Prevalence is the proportion of a population who have the disease at a given time (point prevalence) or at any time during a period (period prevalence). It describes the stock of existing cases and is the measure for planning services, since everyone with the disease needs care, however long ago they became ill.
The proportion of NHANES adults with hypertension is a prevalence. Prevalence depends on both how often people become ill and how long they stay ill. A disease that is common but brief, such as influenza, has a high incidence and, outside an epidemic, a low prevalence. A chronic disease with few new cases each year can have a high prevalence, because the cases accumulate.
The steady-state relation
How exactly does prevalence depend on incidence and duration? Here is the idea in one sentence: in a population whose disease level has settled, as many people must enter the diseased state each year as leave it.
The formal version. Suppose that disease-free people become cases at rate $I$ per person-year. Each case lasts for a time with mean $D$ years, ending in recovery (or death, if those who die are replaced). Let the population have $N$ members and let the prevalence be $P$.
- $N(1 - P)$ people are free of the disease, and they enter it at rate $I$.
- $NP$ people are diseased, and if durations are exponential with mean $D$, they leave at rate $1/D$.
In steady state the two flows are equal:
$$I \, N(1 - P) = \frac{NP}{D} \quad\Longrightarrow\quad \frac{P}{1 - P} = I \times D, \qquad P = \frac{ID}{1 + ID}.$$
The prevalence odds equal incidence times mean duration, and for a rare disease, where $1 - P \approx 1$, this gives $P \approx I \times D$.
The exponential durations only make the outflow easy to write down. Whatever the distribution of durations, the steady state still balances. The number of people who have the disease equals the number who become cases per year times the average number of years each remains a case: $NP = I\,N(1 - P) \times D$, the same relation.
The script checks the relation by simulation, in continuous time. It starts 100,000 people disease-free at time 0 and follows them for 100 years. They become cases at $I = 0.02$ per year and recover after exponential durations with mean $D = 5$ years.
The formula gives $ID = 0.1$ and $P = 0.1/1.1 = 0.09091$, and across 30 seeds the simulated prevalence at year 100 had median 0.09124 (interquartile range 0.09054 to 0.09197). The incidence measured among the disease-free in years 50 to 100 had median 0.02000 per person-year.
Why prevalent cases are long cases
The same simulation shows a less obvious consequence.
Check your understanding. Every episode in the simulation lasts 5 years on average. Take the cases present at year 100 and measure the total length of each one's episode, from start to finish: what will their average be?
About 10 years, not 5: in the simulation it was 10.01 years, twice the mean duration $D = 5$ of all episodes.
This is length-biased sampling. A survey at one moment is more likely to find a long episode than a short one, simply because the long episode occupies more of the time line. The doubling is exact here because two conditions hold. The process is in steady state, with new episodes starting at a constant rate for a long time, and durations are exponential.
Then, at a fixed moment, the time already spent in a current episode is exponential with mean $D$, and so is the time remaining. The second follows from the memoryless property, the first by the same argument run backwards in time, so their sum has mean $2D$.
For other duration distributions the bias remains, but its size changes. A survey at a random moment finds an episode of length $t$ with probability proportional to $t$. So the prevalent episodes have mean length
$$\frac{E(T^2)}{E(T)} = D + \frac{\operatorname{Var}(T)}{D},$$
where $T$ is the duration of an episode. That is twice $D$ when the standard deviation equals the mean, as for the exponential. It is more when durations are more variable, and exactly $D$ only when every episode lasts the same time.
The practical lesson is that a cross-sectional sample of people who have a disease is not a sample of the disease's cases. A prevalence survey over-represents slow, chronic and survivable cases. It under-represents the quickly fatal and the quickly cured. Studies of the causes or the course of a disease are built on incident cases for this reason.
Crude, specific and adjusted rates
Back to Maine and Alaska, whose crude death rates started the article. A crude rate lumps the whole population together, so what does it hide?
A crude rate is computed for a whole population: all deaths in Maine in 2021 over all of Maine's residents. A specific rate is computed for a subgroup.
- An age-specific rate counts the deaths and the population of one age group.
- A sex-specific rate counts those of one sex.
- A cause-specific rate counts only deaths from one cause, with the whole population still at risk.
Rates can be specific for several characteristics at once, such as heart disease deaths among women aged 65 to 74. An adjusted (or standardized) rate is a summary constructed to remove the effect of differences in age or another variable, and is the subject of the next article.
Age-specific rates are the most informative, because age is the strongest single influence on almost every rate of death and disease. The table gives the U.S. figures for 2019, from CDC WONDER. NCHS publishes the same rates (Murphy et al., 2024, Table 6).
| Age | Deaths | Population | Rate per 100,000 | Population share | Share × rate |
|---|---|---|---|---|---|
| Under 1 | 20,921 | 3,783,052 | 553.0 | 0.0115 | 6.4 |
| 1–4 | 3,676 | 15,793,631 | 23.3 | 0.0481 | 1.1 |
| 5–14 | 5,497 | 40,994,163 | 13.4 | 0.1249 | 1.7 |
| 15–24 | 29,771 | 42,687,510 | 69.7 | 0.1300 | 9.1 |
| 25–34 | 59,178 | 45,940,321 | 128.8 | 0.1400 | 18.0 |
| 35–44 | 82,986 | 41,659,144 | 199.2 | 0.1269 | 25.3 |
| 45–54 | 160,393 | 40,874,902 | 392.4 | 0.1245 | 48.9 |
| 55–64 | 374,937 | 42,448,537 | 883.3 | 0.1293 | 114.2 |
| 65–74 | 555,559 | 31,483,433 | 1,764.6 | 0.0959 | 169.3 |
| 75–84 | 688,027 | 15,969,872 | 4,308.3 | 0.0487 | 209.6 |
| 85 and over | 873,746 | 6,604,958 | 13,228.6 | 0.0201 | 266.2 |
| All ages | 2,854,838 | 328,239,523 | 869.7 | 1 | 869.7 |
The rates fall from the first year of life to a minimum at ages 5 to 14. Then they climb steadily, at least doubling with each decade of age from 35 on. The rate at 85 and over is 987 times the rate at 5 to 14. (The all-ages deaths include 147 whose age was not stated.)
The crude rate as a weighted average
The last column of the table hints at how the crude rate is built from the specific ones. Here is the idea in one sentence: the crude rate is an average of the age-specific rates, and each age group counts in proportion to its share of the population.
Let a population be divided into $k$ age groups, with $d_i$ deaths and population $n_i$ in group $i$. Write $r_i = d_i/n_i$ for the age-specific rates, $D = \sum d_i$ for the total deaths and $N = \sum n_i$ for the total population. Then
$$R_{\text{crude}} = \frac{D}{N} = \frac{\sum_i n_i r_i}{N} = \sum_i \frac{n_i}{N}\, r_i.$$
The crude rate is a weighted average of the age-specific rates. The weights are the population's own age distribution, the shares $n_i/N$. This is the grouped-data mean of Measures of Center and Spread, with the age groups as the classes.
As a check, the last column of the table shows the products $(n_i/N)\,r_i$. They add up to 869.70, the crude rate, to rounding; the 147 deaths of unknown age add 0.04.
So two populations can differ in their crude rates for two separate reasons. Their age-specific rates may differ, or their weights, their age structures, may differ. The crude rate cannot tell which.
Maine and Alaska differ sharply in the weights. Maine had the oldest population of the 50 states and the District of Columbia in 2021, measured by the share aged 65 and over (21.7%). Half of its residents were 44 or older. Alaska ranked 48th of the 51 (13.3%), with half its residents aged 35 or younger. And 20.5% of Alaskans were under 15, against 14.9% of Maine's residents.
Check your understanding. Suppose the two states had exactly the same age-specific death rates, those of the whole country in 2021 (Murphy et al., 2024, Table 2). Would their crude rates then be equal?
They would not. Applying those national rates to each state's population gives crude rates of 1,257.9 per 100,000 for Maine and 799.0 for Alaska, a ratio of 1.57, produced by age structure alone.
Now compare the actual crude rates with these predictions: Maine's, 1,258.4, is almost exactly what the national rates predict for a population of its ages. Alaska's, 847.3, is 6% above its prediction. That calculation is the idea behind indirect standardization, developed with confidence intervals in the next article.
The precision of a rate
Before adjusting for age, there is one more question to settle. Could the gap between the two states, or some part of it, be due to chance?
A death rate computed from a complete count of deaths is not subject to sampling error: no sample was drawn. It is still subject to random variation. The number of deaths that occurred is one outcome of a process that could have produced a few more or a few fewer under the same conditions. A comparison between two places, or two years, should not read meaning into differences of that size. NCHS takes this view (Murphy et al., 2024, Technical Notes, "Random variation").
The Poisson model
The usual model treats the number of deaths $D$ as a Poisson random variable with mean $\mu = \lambda N$. Here $\lambda$ is the underlying rate and $N$ the population (or person-years), treated as a fixed constant. The distribution itself is the subject of The Poisson Distribution.
The one property we need is that a Poisson count has variance equal to its mean. The mean is estimated by $D$ itself, so
$$\operatorname{SE}(D) = \sqrt{D}, \qquad \operatorname{SE}(R) = \frac{\sqrt{D}}{N} = \frac{R}{\sqrt{D}}, \qquad \operatorname{RSE}(R) = \frac{\operatorname{SE}(R)}{R} = \frac{1}{\sqrt{D}},$$
where $R = D/N$ is the rate and RSE is the relative standard error.
The Poisson distribution is a good model when each person has a small probability of the event and people die independently of each other. Anderson and Rosenberg (1998, Technical notes) note that it gives "reasonable, conservative" variances for an open population. The binomial model, which assumes a fixed group of people each of whom either dies or does not, tends to understate the variance.
Check your understanding. A city of 10 million and a town of 10,000 each record 100 deaths from some cause. Which rate is the more precise, relative to its size?
Neither: the relative standard error depends only on the number of deaths, not on the population: $1/\sqrt{100} = 10\%$ in both places. The city's rate is a thousand times smaller, and so is its standard error.
For Maine in 2021, $D = 17{,}269$ and $R = 1{,}258.45$ per 100,000, so
$$\operatorname{SE}(R) = \frac{1{,}258.45}{\sqrt{17{,}269}} = 9.58, \qquad \operatorname{RSE} = 0.76\%.$$
The normal 95% confidence interval is $1{,}258.45 \pm 1.96 \times 9.58 = (1{,}239.7,\ 1{,}277.2)$ per 100,000. With seventeen thousand deaths, random variation is a small fraction of the rate.
Exact limits for small counts
What happens when a rate rests on a handful of deaths? The normal approximation fails, as it did for small binomial samples in Confidence Intervals for a Proportion. The Poisson distribution is skewed, and the symmetric interval $D \pm 1.96\sqrt{D}$ can even extend below zero.
Exact limits for the Poisson mean come from the gamma distribution. It is the continuous family to which the Poisson is linked by the identity
$$P(D \ge d \mid \mu) = P(G_d \le \mu),$$
where $G_d$ is a gamma variable with shape $d$ and scale 1. The 95% limits for $\mu$ are
$$L(D) = G^{-1}_{D}(0.025), \qquad U(D) = G^{-1}_{D+1}(0.975).$$
In words, they are the 2.5th percentile of the gamma distribution with shape $D$ and the 97.5th percentile of the one with shape $D + 1$. The limits for the rate are $L(D)/N$ and $U(D)/N$.
These are the formulas NCHS gives, in spreadsheet form, for counts below 100 (Murphy et al., 2024, Technical Notes). They are also the exact Poisson limits that Breslow and Day (1987, §2.3) tabulate. The script reproduces NCHS's table of limit factors $L(D)/D$ and $U(D)/D$ to all six published decimals at the counts it checked (1, 7, 20 and 50 deaths; Table XI).
Suppose a town of 25,000 records 7 deaths from some cause in a year, a rate of 28.0 per 100,000.
- The normal interval is $28.0 \pm 1.96 \times 28.0/\sqrt{7} = (7.3,\ 48.7)$.
- The exact limits for the count are 2.814 and 14.423. So the exact interval for the rate is (11.3, 57.7) per 100,000.
The exact interval is shifted upward and asymmetric about 28.0, as a skewed distribution requires.
Which interval keeps its promise of 95%? Because the count takes whole-number values, the coverage of each interval can be computed exactly for any true mean $\mu$. We add the Poisson probabilities of the counts whose intervals contain $\mu$, as the proportions unit did for the binomial.
For expected counts between 1 and 10, the normal interval's coverage fell as low as 0.632 and averaged 0.889. At $\mu = 5$ it was 0.870, against 0.980 for the exact interval. Between 20 and 100 expected deaths, the normal interval's coverage averaged 0.946, with a low of 0.929. The exact interval's coverage never fell below 0.950 anywhere from 1 to 150. NCHS uses the normal approximation for counts of 100 or more and the gamma limits below that.
Too few deaths to publish
NCHS replaces a crude or age-specific death rate with an asterisk when it is based on fewer than 20 deaths. It describes that threshold as "the equivalent of an RSE of 23% or more" (Murphy et al., 2024, Technical Notes). The equivalence is the formula above: $1/\sqrt{20} = 22.4\%$ and $1/\sqrt{19} = 22.9\%$. So 19 deaths is the largest count whose relative standard error rounds to 23% or more. At 100 deaths the RSE is 10%. A rate based on a handful of deaths can move by half its value from one year to the next without any change in the underlying risk.
Comparing two rates
Now the two states can be compared formally. When both rates rest on 100 or more deaths, NCHS tests their difference with
$$z = \frac{R_1 - R_2}{\sqrt{\operatorname{SE}(R_1)^2 + \operatorname{SE}(R_2)^2}}.$$
This is the same construction as the unpooled comparison of two means or two proportions (Murphy et al., 2024, Technical Notes, formula 12). For Maine and Alaska, $\operatorname{SE}(R_1) = 9.58$ and $\operatorname{SE}(R_2) = 847.3/\sqrt{6{,}208} = 10.75$, so
$$z = \frac{1{,}258.4 - 847.3}{\sqrt{9.58^2 + 10.75^2}} = 28.6.$$
The 95% interval for the difference is (382.9, 439.4) per 100,000.
The rate ratio $R_1/R_2 = 1.485$ is usually analyzed on the log scale. By the delta method,
$$\operatorname{SE}(\log(R_1/R_2)) = \sqrt{\frac{1}{D_1} + \frac{1}{D_2}} = \sqrt{\frac{1}{17{,}269} + \frac{1}{6{,}208}} = 0.0148,$$
and the 95% interval is $1.485 \times e^{\pm 1.96 \times 0.0148} = (1.443,\ 1.529)$.
Check your understanding. A $z$ of 28.6 is far beyond any conventional threshold. Does it show that a Maine resident of a given age was more likely to die than an Alaskan of the same age?
It does not, and this is the misreading the whole unit is built to prevent. The difference is enormous by the standard of random variation: no plausible chance fluctuation produces a $z$ of 28.6. But the test answers only whether the two crude rates differ by more than chance, and they do. It cannot say whether the difference reflects health or demography, because the crude rates have different weights.
The calculation in the previous section already settled part of that. With the same national age-specific rates applied to both states, age structure alone would give Maine a crude rate 57% higher than Alaska's.
Precision and validity are different properties: a large population makes a rate precise, but only a comparison of like with like makes it mean what it appears to.
Summary
| Quantity | Definition |
|---|---|
| Proportion | Part over whole, between 0 and 1: prevalence, risk |
| Ratio | Any quotient of two quantities: sex ratio, infant deaths per live birth |
| Incidence rate | New cases / person-time at risk; units of 1/time |
| Risk over time $t$ (constant rate) | $1 - e^{-\lambda t} \approx \lambda t$ |
| Prevalence and incidence (steady state) | $P/(1 - P) = I \times D$ |
| Crude rate | $D/N = \sum_i (n_i/N)\, r_i$ |
| Standard error of a rate | $R/\sqrt{D}$; relative standard error $1/\sqrt{D}$ |
| Exact 95% limits for a count $D$ | $G^{-1}_{D}(0.025)$ and $G^{-1}_{D+1}(0.975)$ |
| Two rates, 100 or more deaths each | $z = (R_1 - R_2)/\sqrt{\operatorname{SE}_1^2 + \operatorname{SE}_2^2}$; rate ratio with $\operatorname{SE}(\log) = \sqrt{1/D_1 + 1/D_2}$ |
Corrections to the course notes
The unit's notes list prevalence among the types of rate, as a "prevalence rate". The name is common, but prevalence is a proportion of a population at a point or over a period, with no time in its denominator. Keeping the distinction is what makes the relation $P/(1 - P) = I \times D$ intelligible.
The notes give a confidence interval of the form $\text{Rate} \pm 1.96\sqrt{\text{Rate}/\text{Population}} \times \text{Base}$. Read with the rate as a proportion per person (before multiplying by the base), it is the normal interval of this article, $R \pm 1.96\, R/\sqrt{D}$, since $\sqrt{R/N} = R/\sqrt{D}$. It suits rates based on 100 or more events. For fewer, the exact gamma limits above keep their stated coverage.
The notes also give the birth rate as an example of a denominator restricted to women of reproductive age. That restriction defines the general fertility rate (births per 1,000 women aged 15 to 44). The crude birth rate, by convention, divides births by the whole midyear population. The course's formula sheets do not cover this unit.
What we have learned
A rate divides events by the person-time at risk, which is what separates it from a proportion or a ratio. A constant rate implies a risk, $1 - e^{-\lambda t}$, that grows more slowly than the rate times the time. Incidence counts new cases and prevalence counts existing ones. In steady state the prevalence odds equal incidence times mean duration, and a prevalence survey over-represents the long cases. A crude rate is a weighted average of age-specific rates, with the population's own age structure as the weights. Its precision depends only on the number of deaths: a relative standard error of $1/\sqrt{D}$, and exact gamma limits when the count is small.
Maine's crude death rate is far above Alaska's, by a margin no chance fluctuation could produce. Yet the national rates, applied to each state's ages, predict an even larger gap. The remedy is to compare the states' rates at the same ages. Direct standardization applies both states' age-specific rates to a single standard population; indirect standardization applies one set of standard rates to both states and compares their observed deaths with the deaths expected. How do we compare the two states at the same ages, and does the answer depend on whose ages we choose? Direct and Indirect Standardization answers both questions.
In Python
from math import sqrt, exp, log
from scipy import stats
deaths, pop = 17_269, 1_372_247 # Maine, 2021
rate = deaths / pop * 1e5 # 1,258.4 per 100,000
se = rate / sqrt(deaths) # 9.58
print(rate - 1.96 * se, rate + 1.96 * se) # normal 95% interval
d, n = 7, 25_000 # a small count
lo = stats.gamma.ppf(0.025, d) # 2.814
hi = stats.gamma.ppf(0.975, d + 1) # 14.423
print(lo / n * 1e5, hi / n * 1e5) # exact interval: 11.3, 57.7
d1, r1, d2, r2 = 17_269, 1258.4, 6_208, 847.3 # Maine and Alaska
se_log = sqrt(1 / d1 + 1 / d2)
print(r1 / r2, r1 / r2 * exp(-1.96 * se_log), r1 / r2 * exp(1.96 * se_log))
Every number in this article is computed by tools/experiments/applied-statistics/rates-in-public-health.py in the site's repository. The script downloads the Census estimates and queries CDC WONDER, checks both against the published NCHS figures, and runs the simulation with fixed seeds.
References
- Anderson, R. N., and Rosenberg, H. M. (1998). Age standardization of death rates: Implementation of the year 2000 standard. National Vital Statistics Reports, 47(3). National Center for Health Statistics. record · 3 pages
- Breslow, N. E., and Day, N. E. (1987). Statistical Methods in Cancer Research, Volume II: The Design and Analysis of Cohort Studies, Chapter 2 (Rates and rate standardization). IARC Scientific Publications No. 82. International Agency for Research on Cancer. record · 3 pages
- Centers for Disease Control and Prevention, National Center for Health Statistics. Underlying Cause of Death 1999–2020, CDC WONDER online database, from the Multiple Cause of Death Files compiled through the Vital Statistics Cooperative Program. https://wonder.cdc.gov/ucd-icd10.html (queried November 2026). record · 2 pages
- Murphy, S. L., Kochanek, K. D., Xu, J., and Arias, E. (2024). Deaths: Final data for 2021. National Vital Statistics Reports, 73(8). National Center for Health Statistics. Tables 2, 6, 17, IX and XI and the Technical Notes. record · 4 pages
- Pagano, M., Gauvreau, K., and Mattie, H. (2022). Principles of Biostatistics (3rd ed.). CRC Press. The course text. record · 35 pages
- U.S. Census Bureau, Population Division (2022). Annual State Resident Population Estimates for 6 Race Groups (5 Race Alone Groups and Two or More Races) by Age, Sex, and Hispanic Origin: April 1, 2020 to July 1, 2021 (Vintage 2021; released June 2022), file sc-est2021-alldata6. https://www2.census.gov/programs-surveys/popest/datasets/2020-2021/state/asrh/ record · 3 pages